Pre-construction risk assessment of accident scenarios with a hierarchical transformer

The construction industry continues to face persistently high accident rates despite decades of safety regulation. Preventing these accidents depends on knowing which accident scenarios are likely for a given type of work before it begins, yet most data-driven safety studies rely on site imagery or free-text reports that are unavailable at the planning stage. In this study, we present a transformer-based method for pre-construction risk assessment using structured accident records, in which construction type, work process, accident object, and cause are jointly recorded. We derive 8902 instances from the Construction Accident Cases Database of Korea through a documented preprocessing procedure. The model predicts accident objects and causes hierarchically from construction type and work process, capturing the interdependencies among these fields, and assigns each work description a risk score that combines the likelihood of its ranked scenarios, conditional on an accident occurring, with the predicted numbers of fatalities and injuries. Against recurrent, attention-based, and conventional machine learning baselines, the model achieves the highest HIT@1 and NDCG@30 and the lowest fatality and injury errors while matching the strongest baselines on HIT@30 and MAP@30. Risk assessments for representative work descriptions illustrate how the ranked scenarios and scores can inform preventive planning. These findings indicate that structured accident records provide a practical basis for anticipating accident scenarios and their severity before construction work starts.